Masked Grouped Query Attention on T4 GPU Synthetic Performance Benchmark
21.58Performance (TFLOPS)DeepSeek-V3 + Ours
Evaluation Results
| Method | Links | |
|---|---|---|
| DeepSeek-V3 + OursHead Dimension=64, Sequence Length=16k2025.06 | 21.58 | |
| Flex AttentionHead Dimension=128, Sequence Length=8k2025.06 | 15.13 | |
| cuDNNHead Dimension=64, Sequence Length=5122025.06 | 8.21 |